Using AI to detect water bursts helps water companies improve rapid decision-making, be more resource efficient and cut leak response times dramatically. But what makes AI such an effective tool and how does its accuracy and potential keep expanding? Our CTO Neil Edwards explains all.
DETECTING leaks has long been a process involving manual decision-making at key points. A decision to replace a potentially leaky pipe before it bursts; a decision whether an alarm is a leak or just background noise; a decision on which potential leak to investigate first; a decision to backfill a dry hole when a leak can’t be found.
There are two things about this which makes leak detection ripe for the benefits of AI. First, it’s a process. In other words, it’s a series of decisions based on explicit and implicit information.
Second, it requires a lot of human intervention. And if there’s one thing we know about humans, it’s that even the most experienced aren’t always right. Our research suggests that human analysts are about 60 per cent correct. And, it takes us time. None of that is true of AI.
Using AI to detect water bursts
As technology has crept into the leak detection process, more and more data has been available to inform human decisions. Often this data is in multiple proprietary systems which don’t communicate with each other. Analysing it has got harder just as there is more of it to manage.
A digital process is faster, it’s 24/7 and it doesn’t have off days. If you apply some AI knowledge too you add accuracy as well as speed. In the case of true deep learning, AI self-corrects and learns more skills. As with FIDO’s world-first leak-sizing capability, you start to extract ever more valuable information on which to base your response.
In other words, you make better, more informed decisions more quickly.
Faster and more accurate
Put this into a real-life scenario where there are fewer staff on duty. With less resource, there’s less capacity to get the alarms quicker and react to them faster. So, let’s say it’s Easter Sunday. You get 10 alarms and have only three engineering teams. The best outcome would be to reduce the run time of the largest leaks by despatching your engineers to the three biggest.
If you’re lucky, they’re in different DMAs. You could download all the flow graphs from another system, overlay the acoustic data from the loggers and match the points of flow increase to the timings of the logger alarms. In a few hours, you might get one of those large leaks, but you won’t get them all.
You won’t process all 10 and get engineers on the road in under an hour. You might get one scheduled for tomorrow. But by then, you would have lost 24 hours of run time. And for some of those big leaks that’s a phenomenal amount of water.
Reliably better outcomes
An AI-driven response is all about shortening and improving your response times at a basic level. AI eliminates human fallibility from multiple touch points. It processes large volumes of data instantly, assesses outcomes accurately and sends people to where they will be most effective automatically.
FIDO AI uses declarative logic to do this. At its simplest this means that if FIDO spots something happening, it will take a specific course of action. Over time, and the accurate analysis of millions of files, FIDO has learned which are the best outcomes and automates them.
In FIDO’s case, we lifted our clients from a leak/no leak accuracy of 60 per cent to greater than 92per cent with the added ability to rank leaks by size. It gives them a massive performance benefit, with one client reducing their large leak run time from more than 18 days to just three.
Instant mature leak detection
Unlike many AI technologies, FIDO is based on verified data from one of the most advanced water networks in the world. It still has errors, but they’re few. As with humans, errors teach FIDO just as much as getting it right. Arguably more. Knowing where there isn’t a leak is just as valuable to FIDO as knowing where there is one.
FIDO’s original learnt library was based on known outcomes and it continues to grow automatically it amasses more verified results.
No other leak detection AI been able to do this. Operating at arm’s length from their clients, they never see the outcomes of their predictions as we have. One of the joys of this approach is that, having been trained on verified data from advanced networks, FIDO now no longer needs access to confirmatory information. Just an acoustic or vibration file from any hardware device, no matter the make or model.
This gives networks the world over access to FIDO’s mature network solution, with leak sizing, almost overnight, even without their own sensors.
FIDO tracks all the data around a leak from the moment first alarm is received to the moment the leak is fixed. You can only really do this type of meaningful end to end reporting if you integrate the AI analysis with the inputs and outcomes in the workflow system.
This integration is where many technologies fall down. They might be quite brilliant in themselves, but you need to come out of one system to grab data from another.
In terms of the impact on the people using the system, that’s a failure because they aren’t getting a direct benefit themselves in their own working day. I believe this is why many people who understand the theoretical benefit of AI remain opposed to it.
It’s one of the reasons we took a ‘ground up’ approach at FIDO. It’s to encourage its enthusiastic adoption among the people whose buy-in you need in order to put it into practice.
Being embedded into the workflow system, FIDO AI now knows the different performance outcomes that emerge from the verified results of its decisions. It adds this new knowledge to the truths it has already got stored in its learning library. As well as meaning that FIDO does better next time, it also knows how to deliver the outcomes its wants.
With FIDO Leak Central, noise trends from loggers can also be analysed to look for potential emerging leak scenarios even before a leak alarm has been triggered. Effectively this means that acoustic logger estates can now be used to go hunting for leaks rather than just responding reactively to alarms. Our clients are using FIDO AI to detect water bursts and prevent them as well.
Neil Edwards will be delivering a joint case study on this subject with Hannah Wardle of United Utilities at this year’s water sector Global AI & Data Usage Innovation Congress on April 27 to 29, 2021. As lead sponsor, FIDO is offering 25% off entry with our discount code. Click below for details. You can also take part in a survey on attitudes to AI in the water sector for the chance to get a free pass.